Forecasting electricity demand in buildings is now more accurate with Group Encoding (GE), a new method that uses only existing device operation data. Developed by researchers at the Institute of Science Tokyo, the method improved prediction accuracy by 74% in real-world tests. By simplifying high-dimensional binary data, GE supports efficient energy device management, cost reduction, and seamless integration of renewable energy in distributed systems, making it a practical tool for smart energy operation.
As renewable energy becomes more widely adopted, buildings are increasingly outfitted with solar panels, fuel cells, batteries, and other distributed energy systems (DES) to meet their electricity needs. These systems offer significant potential to reduce carbon emissions and improve energy resilience. However, to make full use of these technologies, energy demand must be predicted with high accuracy. Without accurate forecasting, it becomes difficult to balance electricity supply and demand, leading to instability in the power grid, reduced efficiency, and increased costs.
To address this challenge, researchers from the Ihara-Manzhos Laboratory at the Institute of Science Tokyo (Science Tokyo), Japan, have developed a new method called Group Encoding (GE). This technique forecasts a building's electricity demand using only the On/Off status of devices—data that is already collected by most building energy management systems (BEMS). The findings were made available online in Applied Energy on May 5, 2025, and will be published in Volume 392 on August 15, 2025, in a study led by Professor Manabu Ihara and Associate Professor Sergei Manzhos, with contributions from PhD candidate Hyojae Lee and Assistant Professor Keisuke Kameda.
Energy demand forecasting relies on data from BEMS, which tracks various parameters like power generation, consumption, AC settings, and indoor climate conditions. However, each variable adds complexity to the dataset, making analysis more challenging. In contrast, using only the binary On/Off status of devices—data that is already collected by most BEMS—greatly simplifies the dataset while still retaining the key information needed for the control.
"Thanks to the spread of the Internet of Things, On/Off status data, which are minimum information to control devices, can now be collected from building systems easily and at scale. If accurate forecasts can be made using this binary data alone, we can eliminate the need for additional costly sensors," says Ihara.
In the GE method, On/Off data is first gathered from BEMS and tagged according to equipment type. Devices with similar functions—such as heaters, pumps, or air systems—are grouped together. Each device is assigned a weight, either equal or based on its contribution to overall energy use. These weighted signals are then combined to create a single value for each group. The final group of values is input into a machine learning model, which is trained to predict electricity demand.
The researchers tested their method using real-world data from the Environmental Energy Innovation Building at Science Tokyo, which records over 4,000 data points per second or minute, including 1,505 On/Off signals from various energy systems. The team tested the method across four seasonal periods (July 2019, February 2020, July 2021, and February 2022), using one-minute intervals to simulate the rapid fluctuations in electricity demand that are typical during peak summer and winter months.
Compared to conventional label encoding, GE improved forecasting accuracy by 74% in terms of mean absolute error for one-minute-ahead forecasts. For 60-minute-ahead forecasts, GE achieved a mean absolute percentage error of 3.27% and a coefficient of variation of root mean square error of 5.40%, setting a new benchmark for single-building DES forecasting performance. Such a low-cost, high-accuracy forecasting technique could significantly enhance the ability of DES to balance electricity loads, trade in the electricity market, and integrate with variable renewable power sources.
Building on these results, the team is now working to bring the GE method into real-world applications. "We are incorporating this technique into Ene-Swallow®︎, a next-generation intelligent energy management system that controls an advanced Carbon Air Secondary Battery System," says Ihara. The team also has a plan to launch a startup to support the deployment of these technologies and accelerate the integration of variable renewable power sources into the power grid.
By making energy forecasting simpler, cheaper, and more accurate, GE offers a practical solution to one of the major hurdles facing distributed renewable energy systems. As buildings become smarter and more connected, this innovation could play a key role in optimizing energy use, reducing emissions, and ensuring a stable supply of clean electricity.
